Compass: Predicting Biological Activities from Molecular Surface Properties. Performance Comparisons on a Steroid Benchmark

Abstract

We describe a new method, Compass, for predicting the biological activities of molecules based on the activities and three-dimensional structures of other molecules. The method improves on previous techniques by representing only the surface of molecules, by incorporating a nonlinear statistical method, and by automatically choosing conformations and alignments of molecules. We use a benchmark problem of steroid binding affinity prediction to compare the performance of the method with that of two previous systems: CoMFA and a molecular similarity method. Compass predicts steroid affinities substantially more accurately than the others, which represent the state of the art. We present experiments showing that the improved performance depends on each of the technical innovations. IntroductionDrug discovery proceeds largely by trial and error. Typically, thousands of compounds are synthesized for each that finally becomes a drug. Each synthesis costs,

Surflex-Dock 2.1: Robust performance from ligand energetic modeling, ring flexibility, and knowledge-based search

“…Rather than taking the precise pose from a crystal structure, the approach is to find the nearest local optimum and define the score at that optimum as the score for the ligand. This follows the approach developed for Compass, which established the conceptual framework for this approach, termed multiple instance learning within the computational machine learning field [36][37][38][39]. The scoring function was tuned to predict the binding affinities of 34 protein/ligand complexes (overlapping significantly with the Bohm training set), with its output being represented in units of -log(K d ) [2].…”

Untitled

“…Such relational knowledge descriptions are known within the drug design literature as pharmacophores. Successful applications of a machine learning technique to problems related to pharmacophore discovery have been discussed previously (Jain et al, 1994a, Jain et al, 1994b. The domain expert (first author) in the present study suggested a more explicit representation for pharmacophores (Section 2.2).…”

“…The compass algorithm (Jain et al, 1994a, Jain et al, 1994b) overcomes these problems by using a more sophisticated representation of molecular shape, neural network learning methods and adaptation of the alignments. The models produced can be used to predict the activity of new molecules, and, again, visual representations of the results can aid compound design.…”

Determining Molecular Similarity for Drug Discovery using the Wavelet Riemannian Metric

“…Other similarity measures include root-mean-square error on 3D alignment of structures [13], constrained histogram intersection of property distributions on molecular surfaces [22] and Feature Trees [20] which lie between bit string descriptors and 3D descriptors, have also been designed. In Feature Trees, the connectivity of hydrophobic fragments and functional groups in a molecule is represented as a tree and similarity is defined through the match of sub-trees.…”